A Missing-Data Policy Is a Governance Decision

Missing data are often handed to the statistician near the end of a study as if they were only a software problem. By then, the institution may have lost the opportunity to prevent missingness, retain the reason a value is absent, or collect outcomes after a participant stops treatment. The remaining analysis can quantify uncertainty, but it cannot recreate information that governance allowed the workflow to discard.

A missing-data policy should begin before collection and define decision rights across the study lifecycle. Its purpose is not to impose one percentage threshold or one imputation method on every project. It is to ensure that missingness capable of changing a conclusion receives visible, proportionate review.

01Prevention belongs in study design

ICH E9(R1) recommends planning to collect informative reasons for missing data and distinguishing missingness from intercurrent events such as treatment discontinuation, rescue therapy, or death. That distinction affects both what data should continue to be collected and what treatment effect the study estimates.

Governance should therefore require each protocol or data-management plan to identify critical variables, likely causes of absence, steps to reduce avoidable loss, and procedures for follow-up. Burden should be considered: a bloated schedule can itself produce missingness. The team should also determine whether outcomes will still be collected after intervention discontinuation when that is ethically and operationally appropriate.

02Monitor patterns, not a single percentage

An overall 5% missing rate can conceal a severe problem if most missing values come from one site, one participant group, or the primary outcome at the key time point. Conversely, a larger proportion may be less threatening for a secondary variable not used in the main conclusion. A policy should require summaries by variable, site, time, arm where appropriate, and relevant participant characteristics.

Escalation triggers should combine consequence and pattern. Examples include a sudden increase after a form change; imbalance between treatment groups; missingness related to prognosis; absence concentrated in the primary outcome; unexplained generic blanks; or loss that exceeds the assumptions used in the sample-size calculation. These are review triggers, not automatic proof of bias.

03Require a missing-data decision record

For each material issue, the record should state:

1. what is missing and at which time points;
2. the observed pattern and known operational reasons;
3. prevention or recovery actions attempted;
4. the primary analysis assumption and why it is credible;
5. planned sensitivity analyses;
6. who reviewed and approved the response; and
7. how the limitation will be communicated.

This record links operational facts to statistical choices. It also reduces the risk that an attractive method is selected after seeing which analysis gives the preferred answer.

04Make sensitivity analysis proportional to the claim

No statistical technique makes missing data harmless. Multiple imputation, likelihood-based models, weighting, and pattern-mixture approaches each rely on assumptions. The governance question is whether the conclusion remains credible under alternative assumptions that are scientifically plausible.

Sensitivity analysis is especially important when the missingness mechanism cannot be verified, which is common. ICH E9(R1) frames sensitivity analysis as a way to explore robustness of the main estimator to deviations from assumptions. If reasonable scenarios produce meaningfully different conclusions, leaders should not search for a single reassuring result. They should narrow the claim, collect more information if possible, or explicitly state that the evidence is inconclusive.

05Define who can accept residual uncertainty

The analyst can explain assumptions and results, but should not alone decide whether the remaining uncertainty is acceptable for the study’s purpose. The principal investigator, subject-matter expert, statistician, data manager, and—when relevant—ethics, safety, or regulatory representatives bring different responsibilities. High-consequence decisions should be documented with the people who authorized them.

CONSORT 2025 expects transparent reporting of missing-data handling. Institutional governance should go further by making the underlying decision process inspectable before the manuscript is drafted. A reviewer should be able to see not only which method was used, but why collection failed, which alternatives were evaluated, and whether the conclusion survived them.

The most useful first step is small: create a one-page missing-data decision record and require it for every primary outcome with material absence. That transforms missingness from an end-stage technical surprise into a managed research risk.

Review the record again before publication. If the manuscript’s limitations, participant-flow diagram, and analysis description do not match the documented pattern and sensitivity results, the governance process has not yet reached the public research record.

This article is educational and intended for research purposes. It does not provide individual medical advice, diagnosis, or treatment. No patient data were used.